1. RNA-Seq analysis in R (Bioconductor)
Soumya Luthra

2. Analyse the resulting short-read sequences
Index
What is RNA-Seq?
Analysis Pipeline
Mapping
Quantification
Normalization
Differential Expression
Visualization
Demo
EdgeR
DESeq2
Design Experiment
Carefully design the experiment
Purify RNA
Isolate and purify input RNA
Prepare Library
Convert the RNA to cDNA and add sequencing adapters
Sequence
Sequence cDNAs using one of the available NGS platforms
Analysis
Analyse the resulting short-read sequences

3. What is RNA-Seq?
RNA-Seq is the process of sequencing the transcriptome which includes protein coding and non-coding transcripts.
Applications:
Gene (exon, isoform) expression estimation
Differential gene (exon, isoform) expression analysis
Transcriptome assembly - Map exon, intron boundaries, splice junctions
Discovery of novel transcribed regions
Analyse alternate splicing

4. Analysis Pipeline
Slide 19: Vall d’hebron

5. Step 1: Mapping
Align reads against a set of reference genome or transcriptome
Challenges:
Computationally intensive – large number of reads
Mapping reads across splice junctions.
TopHat2 – Splice Aware aligner
Extracting the transcript sequences and using Bowtie to align reads to the virtual transcriptome first
The reads that do not fully map to the transcriptome will then be mapped on the genome and spliced alignment is attempted

6. Step 2: Quantification Identify the read that uniquely map to a gene.
How to resolve overlaps?
Identify the read that uniquely map to a gene.
Tools:
HTSeq-count – Python package
featureCounts() - from Rsubread R package
summarizeOverlaps() - from GenomicsAlignments R package
Counting rules:
Count mapped reads, not base-pairs
Count each read at most once
Discard a read if
It cannot be assigned to any feature
It cannot be uniquely mapped
It can be assigned to more than one gene (ambiguous)
The mates do not map to the same gene
Do not discard if there are read duplicates

7. Step 3: Normalization
Read counts need to be properly normalized to accommodate for the following biases and extract meaningful expression estimates:
Sequencing depth – Higher the sequencing depths, higher the counts
Gene length - Longer transcripts are expected to generate more reads
Count distribution
The main biases that must be accounted for in the normalization and/or differential expression calculations are:

8. Step 4: Differential Expression Analysis
How do the expression levels differ across several conditions?
Challenges:
Count data is discrete – no normal distribution. Cannot perform t-test.
Small number of replicates – can not use permutation methods
Account for variability in measurements across biological replicates of an experiment

9. Poisson Distribution? Mean = Variance
Is read count data Poisson Distributed?
Over-dispersion - variance in RNA-Seq measurements of gene expression are larger than the theoretical values

10. Negative Binomial Distribution
NB has been shown to be a good fit to RNA-Seq data
It is flexible enough to account for biological variability
Model:
Makes the assumption that an observation say Ygj (observed number of reads for gene g sample j, has a mean μgj and a variance of μgj + Φg μ2, where Φg represents over-dispersion relative to poisson distribution.
The mean parameter depends on the sequencing depth as well as on the mount of RNA from gene in the sample
Obtaining good estimates of each gene’s dispersion is critical for statistical testing.
Tools:
EdgeR and DESeq test model the count data using a Negative Binomial distribution and perform statistical tests for differential expression.

11. edgeR
EdgeR treats the Poisson variance as simple sampling variance, and refers to the dispersion estimate as the "biological coefficient of variation.”
Estimating dispersion:
EdgeR shares information across genes to determine a common dispersion. It then calculates a dispersion estimate per gene and shrinks it towards the common dispersion. The gene-specific (referred to in edgeR as tagwise) dispersion estimates are used in the test for differential expression.
Statistical Test:
Simple design - Fischer’s exact test.
Complex design - Generalized linear model framework

12. DESeq
Differential gene expression from count data based on negative binomial distribution.
Offers two transformations for stabilizing the variance of count data
VST – Variance stabilizing Transformation
Regularized log (rlog)

13. Thank You